13 Aug 2025

AI News Digest

🤖 AI-curated 8 stories

Today's Summary

Sam Altman is throwing his weight behind Merge Labs, a new player in the brain-computer interface game, hoping to outsmart Neuralink with a hefty $250M fundraising goal. This move amps up the competition in the neurotech space, where deep pockets and AI prowess are setting the stage for next-gen human-machine hookups. Meanwhile, Microsoft is rolling out GPT-5 across its Copilot tools, making it easier for developers to harness AI’s power, just as Replit’s ‘vibe coding’ is letting solo creators whip up apps in no time, fueling a wave of one-person startups.

Stories

Sam Altman backs Merge Labs to challenge Neuralink with high‑bandwidth brain‑computer interface push

Financial Times reports that Sam Altman is backing a new startup, Merge Labs, which aims to develop high‑bandwidth brain‑computer interfaces and compete with Elon Musk’s Neuralink. The company is reported to be targeting a large fundraising round (reports cite plans to raise roughly $250M) and lists prominent partners on the project. If realized, Merge Labs would raise the stakes in the BCI race — drawing more capital and talent into an area where deep pockets, specialized hardware and AI integration matter. The move highlights growing investor confidence in next‑generation human‑machine interfaces and intensifies direct competition between well‑funded private efforts in the neurotech space.

Refold AI emerges from stealth with $6.5M seed to automate enterprise API integrations

The Economic Times reports that Refold AI, a Bengaluru and San Mateo startup, came out of stealth on Aug. 13, 2025 with a $6.5 million seed round co‑led by Eniac Ventures and Tidal Ventures. Refold offers AI‑driven autonomous agents that learn and maintain API integrations (ERP↔CRM, finance automation, supply‑chain workflows), and already serves over 30 enterprise customers with ~1,500 active users and 30M API calls per month. The funding will expand engineering and integration capabilities. The story underscores continued investor appetite for vertical AI infrastructure that reduces costly manual integration work across large enterprises.

‘Deep ignorance’: filtering training data reduces AI biorisk without wrecking capability

New research from the UK’s AI Security Institute together with EleutherAI shows that selectively removing hazardous domain content from training corpora — a technique the authors call “deep ignorance” — can substantially weaken an open-source model’s ability to generate or help design biological hazards while leaving general capabilities nearly intact. The team trained variants of Pythia‑6.9B on filtered vs. unfiltered datasets and report improved safety on biorisk benchmarks with only minor drops on broader biology tests and negligible compute overhead (~<1% extra). Why it matters: this challenges the widely repeated claim that only ever-larger, unfiltered datasets can produce useful models and suggests an actionable, auditable data‑level safety lever that regulators and researchers could adopt to reduce specific misuse risks without forfeiting model utility.

Position paper argues the centralized AI conference model is unsustainable — proposes a Community‑Federated Conference

An arXiv preprint (submitted Aug 6, revised Aug 12, 2025) presents a data‑driven critique of the current centralized AI conference ecosystem, documenting scientific, environmental, mental‑health and logistical stresses (e.g., rising per‑author publication rates, high carbon footprints, negative online sentiment and venue capacity limits). The authors propose a Community‑Federated Conference (CFC) model that decouples peer review, presentation, and networking into locally organized but globally coordinated components to improve inclusivity, reduce emissions, and relieve pressure on reviewers and venues. Why it matters: the proposal targets systemic, disciplinary‑level change in how AI research is evaluated and disseminated — a shift that could reshape peer review, conference carbon footprints, and community norms if adopted or piloted by major venues (NeurIPS/ICML/ICLR/IJCAI/etc.).

Perplexity shocks the web with a $34.5B unsolicited bid for Google’s Chrome

AI search startup Perplexity submitted an unsolicited all-cash offer of roughly $34.5 billion to buy Google’s Chrome browser (reporting Aug 12, 2025). The move — far larger than Perplexity’s own reported valuation — appears aimed at securing Chrome’s massive user base to accelerate Perplexity’s browser and search ambitions and to capitalize on regulatory pressure that has put Google’s dominance under scrutiny. Even if unlikely to succeed, the bid signals aggressive, headline-grabbing tactics by well‑funded AI rivals and further escalates competitive and antitrust narratives around control of key consumer-facing distribution channels for AI services.

CoreWeave posts blockbuster revenue but widening losses spook investors

CoreWeave reported Q2 results (Aug 12, 2025) showing revenue of about $1.21 billion — well above estimates — driven by strong AI compute demand, but a much larger-than-expected net loss (roughly $290.5M) sent shares lower in after-hours trading. Management raised full‑year revenue guidance even as operating expenses and stock‑based comp ballooned. The report highlights the tension in AI infrastructure names: explosive top-line demand for GPU capacity but heavy capital intensity, customer concentration and margin pressure, raising questions about how quickly such providers can translate growth into sustainable profits.

Microsoft folds GPT‑5 into Copilot — smart model routing and deeper coding help arrive in August

Microsoft has integrated OpenAI’s newly released GPT‑5 across its Copilot ecosystem (Microsoft 365 Copilot, GitHub Copilot, Azure AI Foundry and Copilot Studio), and introduced a “smart mode” that routes tasks to the best model automatically. The change brings GPT‑5’s longer context, stronger reasoning and improved code generation into everyday productivity and developer tools — meaning faster, more capable code completion, enhanced meeting/notes workflows, and easier agent-building for enterprises. For developers and teams this reduces friction when using multiple models, and for learners it makes Copilot a more powerful hands‑on coding tutor. The rollout began in early August and many features (Copilot Pages, mobile voice interactions, expanded grounding on Loop/Teams) are arriving throughout August. URL: https://www.theverge.com/news/753984/microsoft-copilot-gpt-5-model-update

Replit’s ‘vibe coding’ claims: build an app in an afternoon — and a new wave of solo startups

Replit’s CEO says AI‑driven ‘vibe coding’ — using natural‑language prompts + Replit Agent to generate multi‑file apps and live previews — has lowered the barrier to building software, enabling non‑technical creators to ship prototypes and small businesses to launch products quickly. The trend is powering a surge in one‑person startups and offers a practical learning pathway for people who want to learn coding by doing: spawn an app, inspect/edit the generated code, iterate and deploy. Replit’s Agent (including Agent v2 improvements) and growing ecosystem are positioned as both productivity tools and informal, hands‑on learning environments for new coders. URL: https://www.businessinsider.com/replit-ceo-vibe-coding-solo-startups-2025-8